You are on page 1of 7

2003 Nature Publishing Group http://www.nature.



Synchrony-dependent propagation of firing rate in

iteratively constructed networks in vitro
Alex D Reyes
The precise role of synchronous neuronal firing in signal encoding remains unclear. To examine what kinds of signals can be
carried by synchrony, I reproduced a multilayer feedforward network of neurons in an in vitro slice preparation of rat cortex
using an iterative procedure. When constant and time-varying frequency signals were delivered to the network, the firing of
neurons in successive layers became progressively more synchronous. Notably, synchrony in the in vitro network developed
even with uncorrelated input, persisted under a wide range of physiological conditions and was crucial for the stable
propagation of rate signals. The firing rate was represented by a classical rate code in the initial layers, but switched to a
synchrony-based code in the deeper layers.

There has been some debate regarding the role of synchrony in nervous system signaling19. One view is that synchrony is involved in
transmitting temporally precise signals. Under this coding scheme, the
activities of neurons become correlated during a task or during sensory stimulation1023. Another view is that signals are propagated by a
rate code such that the number, and not the timing, of action potentials (APs) is the important variable2,4,2426. In this view, synchrony has
only a minor role in, and may even compromise, rate coding4,26,27.
Detailed analyses of signal propagation have thus far been limited
to models of feedforward networks4,2023,2527. Simulations indicate
that depending on the condition, the network can provide a substrate
for either a temporal or rate code. In a network of integrate-and-fire
neurons, a packet of synchronized APs delivered to the first layer will
propagate through the network2023. In this mode, the network preferentially transmits temporally precise signals: the firing of neurons
in successive layers will either synchronize in the submillisecond
range if the input APs are sufficiently correlated, or the firing will fade
rapidly2023,27. Under certain conditions, synchrony can be eliminated in the presence of background noise that presumably mimics
in vivo conditions4,25. In this mode, information about input rate,
rather than timing, is preferentially propagated through the network.
A caveat with these analyses is that the models make simplifying
assumptions about the biophysical properties of neurons. The degree
to which either temporal or rate coding takes place will likely depend
on the filtering characteristics imparted to the neuron by various
voltage- and time-dependent conductances.
This study uses an iterative procedure to construct an in vitro network of cortical neurons. To examine the coding schemes developed
with models, the iteratively constructed network (ICN) was configured
to have a feedforward architecture. Although considerably simpler than
cortical networks, analyses of feedforward networks have nevertheless
provided useful insights into signal propagation. The experiments tests
specifically whether input rate can be propagated through the network.

To this end, the conditions for the development and elimination of synchrony were examined systematically. The results show that in a simple
feedforward network, synchrony developed under a wide range of
stimulus conditions and network configurations. Synchrony therefore
seems to be the default state. Furthermore, under the conditions of the
experiments, synchrony did not encode temporally precise inputs but
instead was critically involved in ensuring that rate signals propagated
in a stable manner across layers.
Iteratively constructed networks
The ICN consisted of m layers, each consisting of w cells (Fig. 1a). The
network was sparsely connected: each neuron was innervated by 10%
of neurons from the previous layer. Whole-cell recording was established in a pyramidal neuron (in layer 2) using an in vitro slice preparation of rat somatosensory cortex2831. This neuron was driven with
inputs from a population of simulated layer-1 neurons (filled circles
in Fig. 1a)28. The simulated neurons were made to fire repetitively and
asynchronously with respect to each other. Each action potential (AP,
Fig. 1b) resulted in a postsynaptic current (PSC). A subset (10%) of
these neurons was chosen randomly, and their PSC trains were
summed. The resultant waveform (Fig. 1b, Sum1,1) represented the
net synaptic current that the layer-1 neurons combined activity
would generate in a layer-2 neuron. This waveform was then injected
under current clamp (unless otherwise stated) into the pyramidal
neuron to evoke repetitive firing (Fig. 1c, AP2,1). To replicate the
activity of another cell (dashed circles) in layer 2, a new current trace
(Sum1,2) was calculated and injected into the pyramidal neuron to
once again evoke repetitive firing (AP2,2). To complete the first iteration, this process was performed w times, once for each neuron in
layer 2. To propagate the signal into the third layer, the procedure was
repeated except that simulated layer-1 AP trains were replaced with
the evoked layer-2 AP trains. When the summed current was injected

Center for Neural Science, New York University, 4 Washington Place, New York, New York 10003, USA. Correspondence should be addressed to A.D.R.




2003 Nature Publishing Group

Figure 1 Constructing feedforward networks in vitro. (a) Networks had
m layers with w neurons per layer. (b) The firing (AP) of each neuron in the
first layer () was simulated, and associated trains of postsynaptic currents
(PSCs) were calculated. A randomly chosen subset of PSC trains were
summed (Sum1,1). (c) The summed current was injected into the recorded
neuron (which represents a neuron in layer 2) to evoke repetitive action
potentials (AP2,1). A new summed current trace (Sum1,2 in b) was then
calculated from a different subset of PSC trains and again injected into the
neuron. The resultant firing was now equivalent to that of another cell in
layer 2 (AP2,2). This process was performed w times to complete the first
iteration. These evoked APs were then used to calculate a new set of PSC
trains (PSC2,1 to PSC2,w). A subset of these trains was again chosen
randomly, summed, and the current was re-injected back into the cell. Now,
the resultant firing was equivalent to that of a cell in layer 3 (not shown).

into the neuron, the resultant firing was equivalent to that of a cell in
the third layer. The pyramidal neuron was once again stimulated
w times to complete the second iteration. This procedure was iterated
m times, once for each layer. In this manner, the signal introduced in
the first layer could be systematically traced through the network.
The individual PSCs were adjusted so that when injected into a
neuron in vitro, the resultant voltage deflection was comparable to
unitary postsynaptic potentials (PSPs) measured with paired recordings (0.31.0 mV)29,31. To ensure that the evoked firing (Fig. 1c) varied from trial to trial, the amplitudes and latencies of the PSCs were
randomized in such a way that the parameters of the associated PSPs
were within the range of experimentally measured values29,31.
Development of synchrony
In every case examined (n = 60) under a variety of conditions (see
below), neuronal firing was asynchronous for the first 23 layers but
became progressively more synchronous in successive layers. The
occurrences of APs are documented in dot rasters (Fig. 2a), where
each row of dots represents a train of APs in a given cell, and summarized in population stimulus histograms (H in Fig. 2b). In layer 1, the
dots in the raster were distributed uniformly and the histogram was
flat, confirming that the simulated neurons fired asynchronously
with respect to each other. In layer 2, the dots began to cluster, and
the histograms began to develop peaks, indicating that the neurons
had started to fire synchronously. With each passing layer, the clustering and histogram peaks became sharper to approach a steady
level. Magnification of the raster and histogram in layer 11 (boxed


Figure 2 Development of synchrony. (a) Dot rasters documenting the firing of

neurons in a multi-layer network (200 neurons/layer). Each row of dots
represents an action potential (AP) train from a single neuron. (b) Time
histograms (H) and APs for neurons in selected layers. (c) Magnified view of
raster and histogram for boxed regions in a and b. The smooth curve is a
Gaussian fit to the histogram. (d) Plot of normalized cross-correlation
histogram (CCH) peak area versus layer number. Population CCHs (insets)
were compiled by cross-correlating the AP trains of all the cells within a layer
with each other. Solid curve is the best sigmoid fit. Scale bar in b, 20 mV.

regions in Fig. 2a,b) shows that the APs within a cluster were distributed normally (Fig. 2c). The histogram peak was fitted with a
Gaussian function (solid curve), with a standard deviation (s.d.) of
3 ms. To quantify synchrony, cross-correlation histograms (CCHs)
were constructed27,32,33. The average CCHs (calculated from pairwise correlations of all spike trains in each layer) in the initial layers
were flat, whereas those in deeper layers developed peaks that straddle the origin (Fig. 2d, insets). The area of the CCH peak above baseline (normalized by the number of sweeps) increased sigmoidally
with layer. Note that the peak area in networks with bursting neurons
could exceed 1. A sigmoid relation was also obtained with another
metric that treated APs in bursts individually34. The average ( s.d.)
of the Gaussian fits to the CCH peaks past layer 5 was 6.0 3.7 ms
(range 1.613.1; n = 11).
Because the input into layer-2 neurons was calculated from a finite
ensemble of rhythmically firing layer-1 neurons, the injected current
may have had oscillatory temporal structure despite the fact that the
layer-1 neurons were uncorrelated. The resultant transients may have
provided the seed for synchrony in the second layer. To reduce oscillations, a Poisson process was used to determine the AP times of the
layer-1 neurons (Fig. 3a,b). In the same network, synchrony developed whether the layer-1 neurons fired in a Poisson or in a repetitive
manner (Fig. 3c; n = 5). Examination of the AP interval distributions



2003 Nature Publishing Group


Figure 3 Synchrony with Poisson input. (a) Dot rasters of APs in layer 1. The
timing of the APs followed a Poisson process (20 Hz). (b) Interspike interval
distribution of APs in layer 1 (L1) and in layer 2 (L2). (c) Plot of normalized
CCH peak area versus layer when the layer-1 neurons fired rhythmically ()
and Poissonly ().

shows that the firing of neurons in layer 2 did not follow a Poisson
process (Fig. 3b). In general, the firing of cortical neurons in response
to uncorrelated inputs was not Poisson, even in the presence of large
background noise (see below).
Synchrony in modified feedforward networks
To determine whether background excitatory and inhibitory inputs
reduce synchrony4, I incorporated feedforward inhibitory neurons
into the network (Fig. 4a, inset). Excitatory and inhibitory postsynaptic potentials (EPSPs and IPSPs) were evoked in the recorded cell by
using a dynamic clamp circuit to inject current (EPSC and IPSC)3537
(Fig. 4a). Dynamic clamp accurately reproduces the conductance
changes caused by synaptic input. The amplitudes of the EPSPs and
IPSPs were between 0.3 and 1.0 mV2931. Initially, levels of excitatory
and inhibitory inputs were adjusted to be in the balanced configuration so that the mixed input increased the variance but not the mean
of the injected current. When injected, the cells membrane potential
fluctuated in the depolarizing and hyperpolarizing directions. One
problem was that the recorded neuron fired at very low rates. Unlike
in models4,25,26 and in experiments performed under current-clamp
conditions38,39 where voltage fluctuations frequently cross the threshold for APs, the increase in conductance reduced the voltage excursions to the point where APs were rarely evoked. Increasing the
number of inputs had little effect37. In practice, a bias toward excitatory inputs was necessary to generate appreciable firing (>10 Hz;
Fig. 4b) and to prevent the firing rate from fading in successive layers.
Under this condition, adding inhibitory neurons to the network did
not prevent synchrony (n = 5; Fig. 4c).
In the models, the tendency toward synchrony may be due to the
fact that the network is composed of a single cell type. Because the
passive and active properties are the same, the neurons will, on the


Figure 4 Synchrony in networks with excitatory and inhibitory neurons.

(a)Voltage (EPSP, IPSP) traces obtained when currents (EPSC, IPSC) were
injected under dynamic clamp. The network consisted of an equal number
of feedforward excitatory and inhibitory neurons (inset). (b) Firing (upper
trace) caused by injection of summed current (lower trace). (c) Plot of
normalized CCH peak area versus layer for networks with only excitatory
neurons (140 neurons, ) and for networks with both excitatory and
inhibitory neurons (150 each; ). Vertical scale bars: a, 200V, 0.01 nA;
b, 20 mV, 0.25 nA. Horizontal scale bars: a, 100 ms; b, 500 ms.

average, tend to fire at similar times at the stimulus onset (as a result
of identical integration times) and thereafter (as a result of identical
input/output functions). To determine whether synchrony persists in
heterogeneous networks, recordings were made from 24 neurons
simultaneously. The neurons were chosen so that their input resistances and firing responses were different: some fired repetitively with
different degrees of adaptation (Fig. 5a), and others fired repetitive
bursts (Figs. 2 and 3)40,41. The neurons were incorporated into the
network so that each layer had an equal number of each cell type
(Fig. 5a, inset). Synchrony still developed after 23 layers, whether the
network contained two (n = 5), three (n = 4) or four (n = 5) different
cells (not shown).
To further increase the firing variability of each neuron, white
noise (0.1 nA) was added to the summed current and injected
under current, rather than dynamic, clamp (Fig. 5b; n = 7). The noise
evoked a background firing rate of 10 Hz (which increased to 20 Hz
by layer 6) and caused each neuron to fire irregularly (right, distribution of AP intervals). The coefficient of variation of AP intervals was
2.0 in layer 2 and decreased to 1.6 in layer 6. As the dot rasters in layer
6 (Fig. 5c) and the plot of CCH peak area (Fig. 5d, triangles) show,
synchrony was reduced but not eliminated by noise. The effect of
higher level noise was not examined because the coefficients of variation of the AP intervals were already at the high end of those commonly reported in vivo4,42,43. Moreover, the effects of white noise
injection may not accurately mimic in vivo conditions, particularly if
physiological noise arises from uncorrelated inputs39,42 that generate
by a large increase in the neurons conductance (see above).
Synchrony also persisted when: (i) a random steady current bias was
introduced from trial to trial to simulate different initial conditions
(n = 3), (ii) the PSCs showed frequency-dependent changes in
amplitudes30,31,44 (n = 9), (iii) the connection probability between
neurons was less than 10% (range, 1% to 5%; n = 5) or (iv) slow
NMDA-like PSCs were used (n = 6).




2003 Nature Publishing Group

Figure 5 Synchrony in heterogeneous networks. Simultaneous whole-cell
recordings were made from three neurons. The neurons were incorporated
into the network so that each layer effectively had three cell types (inset).
(a) Responses of three neurons to current step injection. (b) Responses of
the same neurons when zero mean white noise current (s.d., 0.1 nA)
was added to the calculated current traces and injected under current
clamp. Arrow marks the start of the stimulus. Histograms on right show
the distribution of AP intervals. (c) Dot rasters of APs in layer 6 in the
presence of noise. The network had 240 neurons per layer. Rasters are
sorted according to the three cell types. (d) Plot of normalized CCH peak
area vs. layer for three different levels of white noise current (given as
s.d.): , 0 nA; , 0.05 nA; , 0.10 nA. Inset shows CCHs in L6.
Vertical scale bars: a. 20mV, 0.25 nA; b, 50 counts; c, 10 counts.
Horizontal scale bars in a and c, 200 ms.

Synchrony-dependent propagation of input rate

To determine whether the network preserves rate signals, the firing frequency of the each simulated layer-1 neuron (input frequency or Fin)
was systematically varied (n = 11). In all layers, the number of APs that
occur during the stimuli increases with Fin (Fig. 6a, upper traces for
both 25 and 55 Hz). The average firing rate of the neurons in each
layer, when calculated over the stimulus interval (1 s) provides information about Fin. Plotting the average firing rate ( s.d.) of neurons in
a given layer (F) versus layer number (L) gives an FL curve. For each
of the six input frequencies, the FL curves reached a steady level by
layers 24 and were distinguishable from each other (Fig. 6b).
The changes in firing rates in the initial and deeper layers occurred
through different mechanisms28,45,46. In layer 2, asynchronous inputs
generated a predominantly steady current that caused the neuron to
fire repetitively (Fig. 6a, lower traces). In this regime, the firing rate
varied with the magnitude of the average current, which in turn varied with input frequency28. In layer 6, synchronous inputs generated
suprathreshold current transients that occurred repetitively. In this
regime, the firing rate was determined by the frequency of the transients. The responses of neurons to asynchronous (or steady current)
and synchronous (or transient current pulses) inputs differ even
when the input rates (or average current) are identical28,46.
Without synchrony, rate signals would propagate in an unstable
manner across layers. This can be seen by calculating the FL curves
that would be obtained if synchrony had not developed in the network. In this hypothetical scenario, the firing of neurons within a
given layer depends only on the rate of arrival of inputs. The firing


Figure 6 Representation and propagation of frequency signals. (a) Action

potentials (upper traces) evoked in layer 2 (L2) and layer 6 (L6) neurons
when input frequencies of 25 and 55 Hz were delivered to the network.
The average current (lower traces) evoked in L2 neurons was steady while
that evoked in L6 neurons contained repetitive transients. (b) Combined
average firing rates ( s.d.) of neurons within a layer plotted against layer
(termed the FL curve) for six input frequencies. (c) FL curves obtained
with 350 (), 500 () and 600 () neurons per layer, Superimposed are
the predicted asynchronous FL curves (dotted; calculated with
Fin = 10 Hz, k1 = 0.02; Pw = 0.1) and the FL curves predicted by a
linear combination of synchronous and asynchronous FL curves (solid).
Curves in iiii correspond respectively to cases where the product
k1*Pw*w was less than, equal to, or greater than 1. (d) FL curve () for
the network in Fig. 2. Superimposed on the FL curve are the predicted
asynchronous () and synchronous () FL curves and the predicted
FL curve obtained by their linear sum (solid curve). Vertical scale bars in
a: 25 mV, 1 nA. Horizontal scale bar: 200 ms.

rate (Fas) evoked by asynchronous input can be approximated in the

unsaturated firing regime by
Fas = k1*Fpre*Npre = k1*Fpre*Pw*w


where k1 is a constant determined experimentally28, Fpre and Npre are

the firing rate and number of presynaptic neurons, and Pw is the connection probability of neurons between layers. The term Pw*Fpre*w
gives the arrival rate of inputs. Because Pw*w is constant, the firing
rate of neurons in the successive layers can be calculated by using
equation (1) iteratively:
Fas(L) = k1*Pw*w*Fas(L1) = (k1*Pw*w)L1*Fin


The predicted asynchronous FL curves decreased when k1*Pw*w < 1

(Fig. 6c, curve i) and increased when k1*Pw*w >1 (curve iii); only
when k1*Pw*w = 1 (curve ii) did the FL curves remain stable. The
experimental FL curves, on the other hand, reached steady-state by
layer 4 (Fig. 6c). In general, the asynchronous FL curves matched the
experimentally measured FL curve only for the first 24 layers.



2003 Nature Publishing Group

To determine whether synchrony-based coding of rate can be generalized to time-varying signals, a sinusoidally modulated Fin was
delivered to the network (Fig. 7; n = 5). As was the case for constant
Fin, neurons started to fire synchronously by layer 3 or 4. Neuronal firing, depicted as the instantaneous firing rate averaged over all neurons within a layer, remained modulated in successive layers (Fig. 7b,
thick lines). The firing rate of neurons in layer 2 varied sinusoidally;
however, those of neurons in deeper layers were somewhat distorted,
suggesting that synchrony-based rate coding may be better suited for
constant or slowly modulated input frequencies.
As with constant Fin, the firing response to time-varying Fin can be
described as a sum of an asynchronous and a synchronous term
(Methods). The predicted asynchrony-based firing rate for the sinusoidal input (dotted lines in Fig. 7b) accurately described firing in
layer 2 but underestimated firing in layer 4 and was nearly zero by
layer 6 (not shown). Adding a term proportional to the level of synchrony improved the predicted firing curve substantially (thin line)
though the curve failed to describe the transient bursts. Nevertheless,
these analyses indicate that without synchrony, time-varying rate signals do not propagate.

Figure 7 Synchrony with sinusoidally modulated inputs. (a) Dot rasters and
histograms of APs in layers 1 and 4. The firing rates of neurons in layer 1
were sinusoidally modulated at 2 Hz with peak/trough values of 45/15 Hz.
(b) APs (first and third traces) and instantaneous firing rates (second and
fourth traces) of neurons in layers 2 and 4. The firing rates predicted from
only the asynchronous component (dotted traces) and from the sum of
asynchronous and synchronous components (thin traces) are superimposed
on the experimentally measured firing rate (thick traces). Vertical scale bar
in a, 25 counts; b, 20 mV.

Synchrony had a crucial role in propagating signals in the deeper

layers. The change in overall firing rate contributed by synchrony (Fs)
can be described by Fs(L) = k2*Fin*AL, where A is the normalized CCH
peak area. For the network in Fig. 6d, the value of k2*Fin was approximated by dividing the firing rate in layer 11 (circles) by the corresponding CCH area (Fig. 2d). By layer 11, the contribution of Fas (Fig. 6d,
diamonds) to the overall firing rate is nearly zero. The synchronous
FL curve (triangles) complemented the asynchronous FL curve in
that it matched the experimental FL curve only past layer 6. This
suggests that firing rate depends on asynchrony in the initial layers
but then on synchrony in the deeper layers.
The experimental FL curve was predicted by linearly combining
the asynchronous and synchronous FL curves (Fig. 6d, solid curve):
Fpred(L) = a*Fas(L) + b*Fs(L)


The coefficients a and b were determined using a least squares fit.

The predicted FL curves were accurate when the asynchronous
FL curve decayed to zero or remained constant (k1*Pw*w 1); they
were less accurate when k1*Pw*w >1 because the asynchronous
FL curve continued to increase while both the experimental and
synchronous FL curves plateaued (Fig. 6c). A better fit might be
obtained if the linear input/output relation in equation (1) is
replaced with a sigmoid function.


Comparison with synchrony in models
In the ICN, as in the models2023,27, the firing of neurons in successive layers either synchronized or degraded to zero. Theoretical
analyses suggest that synchrony develops when a sufficiently large
number of input spikes occur within a time interval (spike
packets). A similar condition likely applies to the ICN, albeit with
input spikes that are distributed uniformly rather than normally.
Unlike in models, synchrony in the ICN was beyond the submillisecond range. Temporal precision was reduced by the variable
PSC latencies incorporated into the ICN, by the tendency of some
neurons to burst, and by the presence of a soft threshold in cortical
neurons that causes APs to occur at variable delays after an EPSP45.
Another difference is that synchrony in the ICN was resistant to
background noise. In integrate-and-fire neurons, unlike in cortical
neurons45, the membrane trajectory in the interspike interval rises
exponentially and spends more time hovering near threshold.
Consequently, relatively smaller voltage fluctuations can cross
threshold and disrupt AP timing. The impact of noise may be
further reduced by the fact that the combined properties of voltage- and time-dependent conductances in cortical neurons
enhance the effectiveness of large, transient inputs (Fig. 6a) for
evoking APs28,38,42,46,47.
In these experiments, synchrony did not convey coincident events
because the input signals were uncorrelated. Rather, synchrony had
an important, albeit permissive, role in propagating rate signals: it
ensured that the repetitive current transients in successive layers
(Fig. 6a) were sufficiently large to evoke APs. Computing rate with the
synchrony-based code requires a time window of at least one interspike interval. For example, to discriminate frequency inputs above
20 Hz, a time window of at least 50 ms is needed. This relatively long
time window suggests that the synchrony-based coding of rate may
apply more to long-lasting events such as the coherent oscillations
that appear in cortex during specific motor-related tasks1316.
However, if the functionally relevant range of firing rates is high, the
time window may not be grossly different from that needed for fast
propagation of rate signals via an asynchrony-based code25. In neurons of the visual cortex, the APs evoked with sine-wave grating stimuli are due to large, synchrony-mediated voltage transients that
occur in the frequency range48,49.


2003 Nature Publishing Group

Comparison with synchrony in vivo
Overall, the level of synchrony in the ICN was higher than
in vivo16,33,50. This discrepancy probably did not result from the use of
a fixed and finite ensemble of AP trains to construct network activity,
as cortical neurons also receive a finite number of inputs. That cortical networks are more complex and therefore generate more noise
also cannot explain the difference. Because voltage fluctuations are
limited by conductance increases that accompany synaptic input, it
seems unlikely that presence of feedback circuits would generate
more variability than noise injected under current clamp (Fig. 5). A
possible explanation is that neurons belonging to a particular network in vivo may multiplex several synchronous signals from other
networks. Any differences in the sets of input signals to each neuron
would decrease synchrony within that network. Indeed, synchronous
barrages may underlie the large firing variability observed in vivo39,42.
At first glance, the high level of synchrony in the ICN seems to be
more symptomatic of epileptiform activity. However, given that the
minimum number of neurons needed to propagate signals (100
300 neurons per layer with Pw = 0.1) is only a small fraction of cortical
neurons, full-scale epileptiform activity need not develop throughout
cortex during normal activity.
Surgical procedures. Surgical, slicing and recording techniques were done as
described previously2831 and followed guidelines established by the NYU
Animal Welfare Committee. Slices were made from the sensorimotor cortex of
Wistar rats (21 days or older). During recordings, slices were perfused with
artificial cerebrospinal fluid (125 mM NaCl, 2.5 mM KCl, 25 mM glucose,
25 mM NaHCO3, 1.25 mM NaH2PO4, 2 mM CaCl2 and 1 mM MgCl2) heated
to 3134 C. Layer-5 pyramidal neurons were identified under infrared differential interference contrast videomicroscopy. Whole-cell recordings were performed with pipettes with d.c. resistances of 510 M when filled with
100 mM potassium gluconate, 20 mM KCl, 4 mM MgATP, 10 mM phosphocreatine, 0.3 mM GTP and 10 mM HEPES). Voltage and current signals were
filtered at 10 kHz and digitized at 210 kHz.
Stimulation protocol. A computer program simulated the activities of a specified number of presynaptic cells. Each simulated cell fired repetitively for 1 s at
a specified average rate (AP, Fig. 1b). To ensure that the neurons firing were
not temporally correlated the following procedures were used: (i) jitter was
added to the interspike intervals (ISIs) such that the ISIs were distributed normally about a mean interval with a standard deviation of 10% of the ISI; (ii)
the start times of the spike trains were uniformly distributed within one ISI.
Each time a simulated cell fired an AP, an associated synaptic current (PSC)
was calculated. The time course of the current was described by PSC(t) =
kamp(1 et/0)et/1 where kamp is the amplitude, and 0 (1 ms) and 1 (2 ms)
are time constants. The individual PSCs were adjusted so that when injected
into a neuron in vitro, the resultant voltage deflection was comparable to unitary postsynaptic potentials (PSPs) measured with paired recordings2931. The
PSCs were convolved with the spike trains of each presynaptic cell.
Each injected current trace (for example, Sum1,1 and Sum1,2) was calculated
by randomly choosing and then summing a subset (110%) of PSC trains at
each layer. The current trace was converted to an analog signal and injected
into the cell via the amplifier and recording electrode. The amplitudes (mean
s.d., 1.0 0.9 mV) and latencies (1.7 0.9 ms) of the PSPs were randomized in
each trial. The number of neurons per layer was adjusted so that when the
summed PSCs were injected, the firing rate of the recorded neuron was equal
to that of the simulated neuron. This ensured that the signal was propagated
successfully from layer to layer. Stimuli were delivered at intervals of 3 s.
For experiments with dynamic clamp, two patch electrodes, one for recording membrane potential and the other for injecting current, were placed at the
soma37. An analog dynamic circuit injected current that was proportional to
the membrane potential (V): PSC(t) = g(t)*(V Erev) where g(t) is the
computer-generated conductance change, and Erev is the reversal potential for
excitatory (0 mV) or inhibitory (80 mV) PSCs. The AP trains of the simu-


lated and recorded neurons were used to generate a set of both excitatory and
inhibitory conductance trains. The two sets of trains were converted to analog
signals and routed through separate channels in the dynamic clamp circuit.
The circuit converted the conductance waveforms to EPSC and IPSC trains
and then summed them before injection into the neuron.
Prediction of temporally modulated firing. The predicted asynchrony- and
synchrony-based firing rates (Fas(t) and Fs(t); Fig. 7) in response to sinusoidal
Fin were calculated as follows. The constants in equations (13) were first
determined by delivering a constant Fin to the network. To calculate Fas(t), the
instantaneous firing rates of the simulated layer-1 and recorded layer-2 neurons were calculated, averaged and plotted against each other for each time
point (Fas(t)2 versus Fas(t)1). The sigmoid fit to the data gives the input/output
function (Fas2 = S(Fas1) of the neuron. This equation was iteratively applied to
determine the firing rate across layers: Fas(t)L = S(Fas(t)L1). Note that equation (1) can also be used to calculate Fas(t)L. The synchrony-based component
is given by: Fs(t)L = k2*AL*F1(t) where the coefficient k2 was determined in
layer 6 (where Fas(t) is zero) by curve fitting and A is the corresponding CCH
peak area. Summing Fas(t) and Fs(t) gives the predicted firing rate in successive
layers (Fig. 7, thin line).
The author wishes to thank H. Cateau, F. Chance, T. Lewis and R. Shapley for
providing helpful comments. This work was supported by National Science
Foundation grant IBN0079619 and by the Edith J. Low-Beer foundation.
The author declares that he has no competing financial interests.
Received 11 March; accepted 9 April 2003
Published online 5 May 2003; doi:10.1038/nn1056
1. Abeles, M. Corticonics 208258 (Cambridge Univ. Press, Cambridge, 1991).
2. Shadlen, M.N. & Newsome, W.T. Noise, neural codes and cortical organization. Curr.
Opin. Neurobiol. 4, 569579 (1994).
3. Ferster, D. & Spruston, N. Cracking the neuronal code. Science 270, 756757
4. Shadlen, M.N. & Newsome, W.T. The variable discharge of cortical neurons: implications for connectivity, computation and information coding. J. Neurosci. 18,
38703896 (1998).
5. Shadlen, M.N. & Movshon, J.A. Synchrony unbound: a critical evaluation of the temporal binding hypothesis. Neuron 24, 6777 (1999).
6. Meister, M. & Berry, M.J. The neural code of the retina. Neuron 22, 435450
7. Borst, A. & Theunissen, F.E. Information theory and neural coding. Nat. Neurosci. 2,
947957 (1999).
8. deCharms, R.C. & Zador, A. Neural representation and the cortical code. Annu. Rev.
Neurosci. 23, 613647 (2000).
9. Van Rullen, R. & Thorpe, S.J. Rate coding versus temporal order coding: what the
retinal ganglion cells tell the visual cortex. Neural Comp. 13, 12551283 (2001).
10. Riehle, A., Grun, S., Diesmann, M. & Aertsen, A. Spike synchronization and rate
modulation differentially involved in motor cortical function. Science 278,
19501953 (1997).
11. Prut, Y. et al. Spatiotemporal structure of cortical activity: properties and behavioral
relevance. J. Neurophysiol. 79, 28572874 (1998).
12. Hatsopoulos, N.G., Ojakangas, C.L., Paninksi, L. & Donoghue, J.P. Information
about movement direction obtained from synchronous activity of motor cortical neurons. Proc. Natl. Acad. Sci. USA 95, 1570615711 (1998).
13. Murthy, V.N. & Fetz, E.E. Coherent 2535 Hz oscillations in the sensorimotor cortex
of awake behaving monkeys. Proc. Natl. Acad. Sci. USA 89, 56705674 (1992).
14. Baker, S.N., Kilner, J.M., Pinches, E.M. & Lemon, R.N. The role of synchrony and
oscillations in the motor output. Exp. Brain Res. 128, 109117 (1999).
15. Conway, B.A. et al., Synchronization between motor cortex and spinal motoneuronal
pool during the performance of a maintained motor task in man. J. Physiol. 489,
917924 (1995).
16. Baker, S.N., Spinks, R., Jackson A. & Lemon, R.N. Synchronization in monkey
motor cortex during a precision grip task. I. Taskdependent modulation in
singleunit synchrony. J. Neurophysiol. 85, 869885 (2001).
17. Feige, B., Aertsen, A. & KristevaFeige, R. Dynamic synchronization between multiple cortical motor areas and muscle activity in phasic voluntary movements.
J. Neurophysiol. 84, 26222629 (2000).
18. Panzeri, S., Petersen, R.S., Schultz, S.R. & Lebedev, M. The role of spike timing in
the coding of stimulus location in rat somatosensory cortex. Neuron, 29, 769777
19. Abeles, M., Bergman, E., Margalit, H. & Vaadia, E. Spatiotemporal firing patterns
in the frontal cortex of behaving monkeys. J. Neurophysiol. 70, 16291638


2003 Nature Publishing Group

20. Diesmann, M., Gewaltig, M.O. & Aertsen, A. Stable propagation of synchronous
spiking in cortical neural networks. Nature 402, 529533 (1999).
21. Cateau, H. & Fukai, T. FokkerPlanck approach to the pulse packet propagation in
synfire chain. Neural Net. 14, 675685 (2001).
22. Diesmann, M., Gewaltig, M., Rotter, S. & Aertsen, A., State space analysis of synchronous spiking in cortical neural networks. Neurocomputing 565, 3840 (2001).
23. Gewaltig M., Diesmann, M. & Aertsen A. Propagation of cortical synfire activity: survival probability in single trials and stability in the mean. Neural Net. 14, 657673
24. Barlow, H.B. Single units and sensation: a neuron doctrine for perceptual psychology? Perception 1, 371394 (1972).
25. van Rossum, M.C., Turrigiano, G.G. & Nelson, S.B. Fast propagation of firing rates
through layered networks of noisy neurons. J. Neurosci. 22, 19561966 (2002).
26. Mazurek, M.E. & Shadlen, M.N. Limits to the temporal fidelity of cortical spike rate
signals. Nat. Neurosci. 5, 463471 (2002).
27. Litvak, V., Sompolinsky, H., Segev, I. & Abeles, M. On the transmission of rate code
in long feed-forward networks with excitatoryinhibitory balance. J. Neurosci. 23,
30063015 (2003).
28. Oviedo, H. & Reyes, A.D. Boosting of neuronal firing evoked with asynchronous and
synchronous inputs to the dendrite. Nat. Neurosci. 5, 261266 (2002).
29. Markram, H., Lubke, J., Frotscher, M., Roth, A. & Sakmann, B. Physiology and
anatomy of synaptic connections between thick tufted pyramidal neurons in the
developing rat neocortex. J. Physiol. 500, 409440 (1997).
30. Reyes, A.D. et al. Target-cellspecific facilitation and depression in neocortical circuits. Nat. Neurosci. 1, 279285 (1998).
31. Reyes, A.D. & Sakmann, B. Developmental switch in the short-term modification of
unitary EPSPs evoked in layer 2/3 and layer 5 pyramidal neurons of rat neocortex.
J. Neurosci. 19, 38273835 (1999).
32. Perkel, D.H., Gerstein, G.L. & Moore, G.P. Neuronal spike trains and stochastic
point processes. II. Simultaneous spike trains. Biophys. J. 7, 419440 (1967).
33. Fetz, E., Toyama, K. & Smith, W. Synaptic interactions between cortical neurons.
Cereb. Cortex 9, 1 (1991).
34. Pinsky, P.F. & Rinzel, J. Synchrony measures for biological neural networks. Biol.
Cybern. 73, 12937 (1995).
35. Sharp, A., ONeil, M.B., Abbott, L.F. & Marder, E. The dynamic clamp: artificial conductances in biological neurons. TINS 16, 389394 (1993).


36. Reyes, A.D., Rubel, E.W. & Spain, W.J. In vitro analysis of optimal stimuli for phaselocking and time-delayed modulation of firing in avian nucleus laminaris neurons.
J. Neurosci. 16, 9931007 (1996).
37. Chance, F., Abbott, L.F. & Reyes, A.D. Gain modulation from background synaptic
input. Neuron 35, 773782 (2002).
38. Mainen, Z.F. & Sejnowski, T.J. Reliability of spike timing in neocortical neurons.
Science 268, 15031506 (1995).
39. Stevens, C.F. & Zador, A. Input synchrony and the irregular firing of cortical neurons.
Nat. Neurosci. 1, 210217 (1998).
40. Connors, B.W. & Gutnick, M.J. Intrinsic firing patterns of diverse neocortical neurons. Trends Neurosci. 13, 99104 (1990).
41. Schwindt, P.C., OBrien, J.A. & Crill, W.E. Quantitative analysis of firing properties of
pyramidal neurons from layer 5 of rat sensorimotor cortex. J. Neurophysiol. 77,
24842498 (1997).
42. Softky, W. & Koch, C. The highly irregular firing of cortical cells is inconsistent with
temporal integration of random EPSPs. J. Neurosci. 13, 334350 (1993).
43. Holt, G.R., Softky, W.R., Koch, C. & Douglas, R.J. Comparison of discharge variability in vitro and in vivo in cat visual cortex neurons. J. Neurophysiol. 75, 18061814
44. Markram, H., Wang, Y. & Tsodyks, M. Differential signaling via the same axon of neocortical pyramidal neurons. Proc. Natl. Acad. Sci. USA 9, 53235328 (1998).
45. Reyes, A.D. & Fetz, E.E. Effects of transient depolarizing potentials on the firing rate
of cat neocortical neurons. J. Neurophysiol. 69, 16731682 (1993).
46. Reyes, A.D. & Fetz, E.E. Two modes of interspike interval shortening by brief transient depolarizations in cat neocortical neurons. J. Neurophysiol. 69, 16611672
47. Baker, S.N. Quantification of the relative efficacies of asynchronous and oscillating inputs to a motoneuron pool using a computer model. J. Physiol. 504, 116
48. Azouz, R. & Gray, C.M. Cellular mechanisms contributing to response variability of
cortical neurons in vivo. J. Neurosci. 19, 22092223 (1999).
49. Azouz, R. & Gray, C.M. Adaptive coincidence detection and dynamic gain control in
visual cortical neurons in vivo. Neuron 37, 513523 (2003).
50. Matsumura, M. et al. Synaptic interactions between primate precentral cortex neurons revealed by spiketriggered averaging of intracellular membrane potentials
in vivo. J. Neurosci. 16, 77577767 (1996).